1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2021 WANG Xuerui <git@xen0n.name>
5 *
6 * Based on tcg/riscv/tcg-target.c.inc
7 *
8 * Copyright (c) 2018 SiFive, Inc
9 * Copyright (c) 2008-2009 Arnaud Patard <arnaud.patard@rtp-net.org>
10 * Copyright (c) 2009 Aurelien Jarno <aurelien@aurel32.net>
11 * Copyright (c) 2008 Fabrice Bellard
12 *
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this software and associated documentation files (the "Software"), to deal
15 * in the Software without restriction, including without limitation the rights
16 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
17 * copies of the Software, and to permit persons to whom the Software is
18 * furnished to do so, subject to the following conditions:
19 *
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
26 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
28 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
29 * THE SOFTWARE.
30 */
31
32#include "../tcg-ldst.c.inc"
33
34#ifdef CONFIG_DEBUG_TCG
35static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
36    "zero",
37    "ra",
38    "tp",
39    "sp",
40    "a0",
41    "a1",
42    "a2",
43    "a3",
44    "a4",
45    "a5",
46    "a6",
47    "a7",
48    "t0",
49    "t1",
50    "t2",
51    "t3",
52    "t4",
53    "t5",
54    "t6",
55    "t7",
56    "t8",
57    "r21", /* reserved in the LP64* ABI, hence no ABI name */
58    "s9",
59    "s0",
60    "s1",
61    "s2",
62    "s3",
63    "s4",
64    "s5",
65    "s6",
66    "s7",
67    "s8"
68};
69#endif
70
71static const int tcg_target_reg_alloc_order[] = {
72    /* Registers preserved across calls */
73    /* TCG_REG_S0 reserved for TCG_AREG0 */
74    TCG_REG_S1,
75    TCG_REG_S2,
76    TCG_REG_S3,
77    TCG_REG_S4,
78    TCG_REG_S5,
79    TCG_REG_S6,
80    TCG_REG_S7,
81    TCG_REG_S8,
82    TCG_REG_S9,
83
84    /* Registers (potentially) clobbered across calls */
85    TCG_REG_T0,
86    TCG_REG_T1,
87    TCG_REG_T2,
88    TCG_REG_T3,
89    TCG_REG_T4,
90    TCG_REG_T5,
91    TCG_REG_T6,
92    TCG_REG_T7,
93    TCG_REG_T8,
94
95    /* Argument registers, opposite order of allocation.  */
96    TCG_REG_A7,
97    TCG_REG_A6,
98    TCG_REG_A5,
99    TCG_REG_A4,
100    TCG_REG_A3,
101    TCG_REG_A2,
102    TCG_REG_A1,
103    TCG_REG_A0,
104};
105
106static const int tcg_target_call_iarg_regs[] = {
107    TCG_REG_A0,
108    TCG_REG_A1,
109    TCG_REG_A2,
110    TCG_REG_A3,
111    TCG_REG_A4,
112    TCG_REG_A5,
113    TCG_REG_A6,
114    TCG_REG_A7,
115};
116
117static const int tcg_target_call_oarg_regs[] = {
118    TCG_REG_A0,
119    TCG_REG_A1,
120};
121
122#ifndef CONFIG_SOFTMMU
123#define USE_GUEST_BASE     (guest_base != 0)
124#define TCG_GUEST_BASE_REG TCG_REG_S1
125#endif
126
127#define TCG_CT_CONST_ZERO  0x100
128#define TCG_CT_CONST_S12   0x200
129#define TCG_CT_CONST_N12   0x400
130#define TCG_CT_CONST_U12   0x800
131#define TCG_CT_CONST_C12   0x1000
132#define TCG_CT_CONST_WSZ   0x2000
133
134#define ALL_GENERAL_REGS      MAKE_64BIT_MASK(0, 32)
135/*
136 * For softmmu, we need to avoid conflicts with the first 5
137 * argument registers to call the helper.  Some of these are
138 * also used for the tlb lookup.
139 */
140#ifdef CONFIG_SOFTMMU
141#define SOFTMMU_RESERVE_REGS  MAKE_64BIT_MASK(TCG_REG_A0, 5)
142#else
143#define SOFTMMU_RESERVE_REGS  0
144#endif
145
146
147static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
148{
149    return sextract64(val, pos, len);
150}
151
152/* test if a constant matches the constraint */
153static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
154{
155    if (ct & TCG_CT_CONST) {
156        return true;
157    }
158    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
159        return true;
160    }
161    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
162        return true;
163    }
164    if ((ct & TCG_CT_CONST_N12) && -val == sextreg(-val, 0, 12)) {
165        return true;
166    }
167    if ((ct & TCG_CT_CONST_U12) && val >= 0 && val <= 0xfff) {
168        return true;
169    }
170    if ((ct & TCG_CT_CONST_C12) && ~val >= 0 && ~val <= 0xfff) {
171        return true;
172    }
173    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
174        return true;
175    }
176    return false;
177}
178
179/*
180 * Relocations
181 */
182
183/*
184 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
185 * complicated; a whopping stack machine is needed to stuff the fields, at
186 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
187 * needed.
188 *
189 * Hence, define our own simpler relocation types. Numbers are chosen as to
190 * not collide with potential future additions to the true ELF relocation
191 * type enum.
192 */
193
194/* Field Sk16, shifted right by 2; suitable for conditional jumps */
195#define R_LOONGARCH_BR_SK16     256
196/* Field Sd10k16, shifted right by 2; suitable for B and BL */
197#define R_LOONGARCH_BR_SD10K16  257
198
199static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
200{
201    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
202    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
203
204    tcg_debug_assert((offset & 3) == 0);
205    offset >>= 2;
206    if (offset == sextreg(offset, 0, 16)) {
207        *src_rw = deposit64(*src_rw, 10, 16, offset);
208        return true;
209    }
210
211    return false;
212}
213
214static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
215                             const tcg_insn_unit *target)
216{
217    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
218    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
219
220    tcg_debug_assert((offset & 3) == 0);
221    offset >>= 2;
222    if (offset == sextreg(offset, 0, 26)) {
223        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
224        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
225        return true;
226    }
227
228    return false;
229}
230
231static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
232                        intptr_t value, intptr_t addend)
233{
234    tcg_debug_assert(addend == 0);
235    switch (type) {
236    case R_LOONGARCH_BR_SK16:
237        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
238    case R_LOONGARCH_BR_SD10K16:
239        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
240    default:
241        g_assert_not_reached();
242    }
243}
244
245#include "tcg-insn-defs.c.inc"
246
247/*
248 * TCG intrinsics
249 */
250
251static void tcg_out_mb(TCGContext *s, TCGArg a0)
252{
253    /* Baseline LoongArch only has the full barrier, unfortunately.  */
254    tcg_out_opc_dbar(s, 0);
255}
256
257static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
258{
259    if (ret == arg) {
260        return true;
261    }
262    switch (type) {
263    case TCG_TYPE_I32:
264    case TCG_TYPE_I64:
265        /*
266         * Conventional register-register move used in LoongArch is
267         * `or dst, src, zero`.
268         */
269        tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
270        break;
271    default:
272        g_assert_not_reached();
273    }
274    return true;
275}
276
277static bool imm_part_needs_loading(bool high_bits_are_ones,
278                                   tcg_target_long part)
279{
280    if (high_bits_are_ones) {
281        return part != -1;
282    } else {
283        return part != 0;
284    }
285}
286
287/* Loads a 32-bit immediate into rd, sign-extended.  */
288static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
289{
290    tcg_target_long lo = sextreg(val, 0, 12);
291    tcg_target_long hi12 = sextreg(val, 12, 20);
292
293    /* Single-instruction cases.  */
294    if (lo == val) {
295        /* val fits in simm12: addi.w rd, zero, val */
296        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
297        return;
298    }
299    if (0x800 <= val && val <= 0xfff) {
300        /* val fits in uimm12: ori rd, zero, val */
301        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
302        return;
303    }
304
305    /* High bits must be set; load with lu12i.w + optional ori.  */
306    tcg_out_opc_lu12i_w(s, rd, hi12);
307    if (lo != 0) {
308        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
309    }
310}
311
312static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
313                         tcg_target_long val)
314{
315    /*
316     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
317     * with dedicated instructions for filling the respective bitfields
318     * below:
319     *
320     *        6                   5                   4               3
321     *  3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2
322     * +-----------------------+---------------------------------------+...
323     * |          hi52         |                  hi32                 |
324     * +-----------------------+---------------------------------------+...
325     *       3                   2                   1
326     *     1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
327     * ...+-------------------------------------+-------------------------+
328     *    |                 hi12                |            lo           |
329     * ...+-------------------------------------+-------------------------+
330     *
331     * Check if val belong to one of the several fast cases, before falling
332     * back to the slow path.
333     */
334
335    intptr_t pc_offset;
336    tcg_target_long val_lo, val_hi, pc_hi, offset_hi;
337    tcg_target_long hi32, hi52;
338    bool rd_high_bits_are_ones;
339
340    /* Value fits in signed i32.  */
341    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
342        tcg_out_movi_i32(s, rd, val);
343        return;
344    }
345
346    /* PC-relative cases.  */
347    pc_offset = tcg_pcrel_diff(s, (void *)val);
348    if (pc_offset == sextreg(pc_offset, 0, 22) && (pc_offset & 3) == 0) {
349        /* Single pcaddu2i.  */
350        tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
351        return;
352    }
353
354    if (pc_offset == (int32_t)pc_offset) {
355        /* Offset within 32 bits; load with pcalau12i + ori.  */
356        val_lo = sextreg(val, 0, 12);
357        val_hi = val >> 12;
358        pc_hi = (val - pc_offset) >> 12;
359        offset_hi = val_hi - pc_hi;
360
361        tcg_debug_assert(offset_hi == sextreg(offset_hi, 0, 20));
362        tcg_out_opc_pcalau12i(s, rd, offset_hi);
363        if (val_lo != 0) {
364            tcg_out_opc_ori(s, rd, rd, val_lo & 0xfff);
365        }
366        return;
367    }
368
369    hi32 = sextreg(val, 32, 20);
370    hi52 = sextreg(val, 52, 12);
371
372    /* Single cu52i.d case.  */
373    if (ctz64(val) >= 52) {
374        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
375        return;
376    }
377
378    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
379    tcg_out_movi_i32(s, rd, val);
380    rd_high_bits_are_ones = (int32_t)val < 0;
381
382    if (imm_part_needs_loading(rd_high_bits_are_ones, hi32)) {
383        tcg_out_opc_cu32i_d(s, rd, hi32);
384        rd_high_bits_are_ones = hi32 < 0;
385    }
386
387    if (imm_part_needs_loading(rd_high_bits_are_ones, hi52)) {
388        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
389    }
390}
391
392static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
393{
394    tcg_out_opc_andi(s, ret, arg, 0xff);
395}
396
397static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
398{
399    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
400}
401
402static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
403{
404    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
405}
406
407static void tcg_out_ext8s(TCGContext *s, TCGReg ret, TCGReg arg)
408{
409    tcg_out_opc_sext_b(s, ret, arg);
410}
411
412static void tcg_out_ext16s(TCGContext *s, TCGReg ret, TCGReg arg)
413{
414    tcg_out_opc_sext_h(s, ret, arg);
415}
416
417static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
418{
419    tcg_out_opc_addi_w(s, ret, arg, 0);
420}
421
422static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
423                           TCGReg a0, TCGReg a1, TCGReg a2,
424                           bool c2, bool is_32bit)
425{
426    if (c2) {
427        /*
428         * Fast path: semantics already satisfied due to constraint and
429         * insn behavior, single instruction is enough.
430         */
431        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
432        /* all clz/ctz insns belong to DJ-format */
433        tcg_out32(s, encode_dj_insn(opc, a0, a1));
434        return;
435    }
436
437    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
438    /* a0 = a1 ? REG_TMP0 : a2 */
439    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
440    tcg_out_opc_masknez(s, a0, a2, a1);
441    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
442}
443
444static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
445                            TCGReg arg1, TCGReg arg2, bool c2)
446{
447    TCGReg tmp;
448
449    if (c2) {
450        tcg_debug_assert(arg2 == 0);
451    }
452
453    switch (cond) {
454    case TCG_COND_EQ:
455        if (c2) {
456            tmp = arg1;
457        } else {
458            tcg_out_opc_sub_d(s, ret, arg1, arg2);
459            tmp = ret;
460        }
461        tcg_out_opc_sltui(s, ret, tmp, 1);
462        break;
463    case TCG_COND_NE:
464        if (c2) {
465            tmp = arg1;
466        } else {
467            tcg_out_opc_sub_d(s, ret, arg1, arg2);
468            tmp = ret;
469        }
470        tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
471        break;
472    case TCG_COND_LT:
473        tcg_out_opc_slt(s, ret, arg1, arg2);
474        break;
475    case TCG_COND_GE:
476        tcg_out_opc_slt(s, ret, arg1, arg2);
477        tcg_out_opc_xori(s, ret, ret, 1);
478        break;
479    case TCG_COND_LE:
480        tcg_out_setcond(s, TCG_COND_GE, ret, arg2, arg1, false);
481        break;
482    case TCG_COND_GT:
483        tcg_out_setcond(s, TCG_COND_LT, ret, arg2, arg1, false);
484        break;
485    case TCG_COND_LTU:
486        tcg_out_opc_sltu(s, ret, arg1, arg2);
487        break;
488    case TCG_COND_GEU:
489        tcg_out_opc_sltu(s, ret, arg1, arg2);
490        tcg_out_opc_xori(s, ret, ret, 1);
491        break;
492    case TCG_COND_LEU:
493        tcg_out_setcond(s, TCG_COND_GEU, ret, arg2, arg1, false);
494        break;
495    case TCG_COND_GTU:
496        tcg_out_setcond(s, TCG_COND_LTU, ret, arg2, arg1, false);
497        break;
498    default:
499        g_assert_not_reached();
500        break;
501    }
502}
503
504/*
505 * Branch helpers
506 */
507
508static const struct {
509    LoongArchInsn op;
510    bool swap;
511} tcg_brcond_to_loongarch[] = {
512    [TCG_COND_EQ] =  { OPC_BEQ,  false },
513    [TCG_COND_NE] =  { OPC_BNE,  false },
514    [TCG_COND_LT] =  { OPC_BGT,  true  },
515    [TCG_COND_GE] =  { OPC_BLE,  true  },
516    [TCG_COND_LE] =  { OPC_BLE,  false },
517    [TCG_COND_GT] =  { OPC_BGT,  false },
518    [TCG_COND_LTU] = { OPC_BGTU, true  },
519    [TCG_COND_GEU] = { OPC_BLEU, true  },
520    [TCG_COND_LEU] = { OPC_BLEU, false },
521    [TCG_COND_GTU] = { OPC_BGTU, false }
522};
523
524static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
525                           TCGReg arg2, TCGLabel *l)
526{
527    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
528
529    tcg_debug_assert(op != 0);
530
531    if (tcg_brcond_to_loongarch[cond].swap) {
532        TCGReg t = arg1;
533        arg1 = arg2;
534        arg2 = t;
535    }
536
537    /* all conditional branch insns belong to DJSk16-format */
538    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
539    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
540}
541
542static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
543{
544    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
545    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
546
547    tcg_debug_assert((offset & 3) == 0);
548    if (offset == sextreg(offset, 0, 28)) {
549        /* short jump: +/- 256MiB */
550        if (tail) {
551            tcg_out_opc_b(s, offset >> 2);
552        } else {
553            tcg_out_opc_bl(s, offset >> 2);
554        }
555    } else if (offset == sextreg(offset, 0, 38)) {
556        /* long jump: +/- 256GiB */
557        tcg_target_long lo = sextreg(offset, 0, 18);
558        tcg_target_long hi = offset - lo;
559        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
560        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
561    } else {
562        /* far jump: 64-bit */
563        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
564        tcg_target_long hi = (tcg_target_long)arg - lo;
565        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
566        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
567    }
568}
569
570static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg)
571{
572    tcg_out_call_int(s, arg, false);
573}
574
575/*
576 * Load/store helpers
577 */
578
579static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
580                         TCGReg addr, intptr_t offset)
581{
582    intptr_t imm12 = sextreg(offset, 0, 12);
583
584    if (offset != imm12) {
585        intptr_t diff = offset - (uintptr_t)s->code_ptr;
586
587        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
588            imm12 = sextreg(diff, 0, 12);
589            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
590        } else {
591            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
592            if (addr != TCG_REG_ZERO) {
593                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
594            }
595        }
596        addr = TCG_REG_TMP2;
597    }
598
599    switch (opc) {
600    case OPC_LD_B:
601    case OPC_LD_BU:
602    case OPC_LD_H:
603    case OPC_LD_HU:
604    case OPC_LD_W:
605    case OPC_LD_WU:
606    case OPC_LD_D:
607    case OPC_ST_B:
608    case OPC_ST_H:
609    case OPC_ST_W:
610    case OPC_ST_D:
611        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
612        break;
613    default:
614        g_assert_not_reached();
615    }
616}
617
618static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
619                       TCGReg arg1, intptr_t arg2)
620{
621    bool is_32bit = type == TCG_TYPE_I32;
622    tcg_out_ldst(s, is_32bit ? OPC_LD_W : OPC_LD_D, arg, arg1, arg2);
623}
624
625static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
626                       TCGReg arg1, intptr_t arg2)
627{
628    bool is_32bit = type == TCG_TYPE_I32;
629    tcg_out_ldst(s, is_32bit ? OPC_ST_W : OPC_ST_D, arg, arg1, arg2);
630}
631
632static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
633                        TCGReg base, intptr_t ofs)
634{
635    if (val == 0) {
636        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
637        return true;
638    }
639    return false;
640}
641
642/*
643 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
644 */
645
646#if defined(CONFIG_SOFTMMU)
647/*
648 * helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
649 *                                     MemOpIdx oi, uintptr_t ra)
650 */
651static void * const qemu_ld_helpers[4] = {
652    [MO_8]  = helper_ret_ldub_mmu,
653    [MO_16] = helper_le_lduw_mmu,
654    [MO_32] = helper_le_ldul_mmu,
655    [MO_64] = helper_le_ldq_mmu,
656};
657
658/*
659 * helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
660 *                                     uintxx_t val, MemOpIdx oi,
661 *                                     uintptr_t ra)
662 */
663static void * const qemu_st_helpers[4] = {
664    [MO_8]  = helper_ret_stb_mmu,
665    [MO_16] = helper_le_stw_mmu,
666    [MO_32] = helper_le_stl_mmu,
667    [MO_64] = helper_le_stq_mmu,
668};
669
670/* We expect to use a 12-bit negative offset from ENV.  */
671QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) > 0);
672QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) < -(1 << 11));
673
674static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
675{
676    tcg_out_opc_b(s, 0);
677    return reloc_br_sd10k16(s->code_ptr - 1, target);
678}
679
680/*
681 * Emits common code for TLB addend lookup, that eventually loads the
682 * addend in TCG_REG_TMP2.
683 */
684static void tcg_out_tlb_load(TCGContext *s, TCGReg addrl, MemOpIdx oi,
685                             tcg_insn_unit **label_ptr, bool is_load)
686{
687    MemOp opc = get_memop(oi);
688    unsigned s_bits = opc & MO_SIZE;
689    unsigned a_bits = get_alignment_bits(opc);
690    tcg_target_long compare_mask;
691    int mem_index = get_mmuidx(oi);
692    int fast_ofs = TLB_MASK_TABLE_OFS(mem_index);
693    int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
694    int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
695
696    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
697    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
698
699    tcg_out_opc_srli_d(s, TCG_REG_TMP2, addrl,
700                    TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
701    tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
702    tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
703
704    /* Load the tlb comparator and the addend.  */
705    tcg_out_ld(s, TCG_TYPE_TL, TCG_REG_TMP0, TCG_REG_TMP2,
706               is_load ? offsetof(CPUTLBEntry, addr_read)
707               : offsetof(CPUTLBEntry, addr_write));
708    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
709               offsetof(CPUTLBEntry, addend));
710
711    /* We don't support unaligned accesses.  */
712    if (a_bits < s_bits) {
713        a_bits = s_bits;
714    }
715    /* Clear the non-page, non-alignment bits from the address.  */
716    compare_mask = (tcg_target_long)TARGET_PAGE_MASK | ((1 << a_bits) - 1);
717    tcg_out_movi(s, TCG_TYPE_TL, TCG_REG_TMP1, compare_mask);
718    tcg_out_opc_and(s, TCG_REG_TMP1, TCG_REG_TMP1, addrl);
719
720    /* Compare masked address with the TLB entry.  */
721    label_ptr[0] = s->code_ptr;
722    tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
723
724    /* TLB Hit - addend in TCG_REG_TMP2, ready for use.  */
725}
726
727static void add_qemu_ldst_label(TCGContext *s, int is_ld, MemOpIdx oi,
728                                TCGType type,
729                                TCGReg datalo, TCGReg addrlo,
730                                void *raddr, tcg_insn_unit **label_ptr)
731{
732    TCGLabelQemuLdst *label = new_ldst_label(s);
733
734    label->is_ld = is_ld;
735    label->oi = oi;
736    label->type = type;
737    label->datalo_reg = datalo;
738    label->datahi_reg = 0; /* unused */
739    label->addrlo_reg = addrlo;
740    label->addrhi_reg = 0; /* unused */
741    label->raddr = tcg_splitwx_to_rx(raddr);
742    label->label_ptr[0] = label_ptr[0];
743}
744
745static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
746{
747    MemOpIdx oi = l->oi;
748    MemOp opc = get_memop(oi);
749    MemOp size = opc & MO_SIZE;
750    TCGType type = l->type;
751
752    /* resolve label address */
753    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
754        return false;
755    }
756
757    /* call load helper */
758    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_A0, TCG_AREG0);
759    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_A1, l->addrlo_reg);
760    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A2, oi);
761    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A3, (tcg_target_long)l->raddr);
762
763    tcg_out_call(s, qemu_ld_helpers[size]);
764
765    switch (opc & MO_SSIZE) {
766    case MO_SB:
767        tcg_out_ext8s(s, l->datalo_reg, TCG_REG_A0);
768        break;
769    case MO_SW:
770        tcg_out_ext16s(s, l->datalo_reg, TCG_REG_A0);
771        break;
772    case MO_SL:
773        tcg_out_ext32s(s, l->datalo_reg, TCG_REG_A0);
774        break;
775    case MO_UL:
776        if (type == TCG_TYPE_I32) {
777            /* MO_UL loads of i32 should be sign-extended too */
778            tcg_out_ext32s(s, l->datalo_reg, TCG_REG_A0);
779            break;
780        }
781        /* fallthrough */
782    default:
783        tcg_out_mov(s, type, l->datalo_reg, TCG_REG_A0);
784        break;
785    }
786
787    return tcg_out_goto(s, l->raddr);
788}
789
790static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
791{
792    MemOpIdx oi = l->oi;
793    MemOp opc = get_memop(oi);
794    MemOp size = opc & MO_SIZE;
795
796    /* resolve label address */
797    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
798        return false;
799    }
800
801    /* call store helper */
802    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_A0, TCG_AREG0);
803    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_A1, l->addrlo_reg);
804    switch (size) {
805    case MO_8:
806        tcg_out_ext8u(s, TCG_REG_A2, l->datalo_reg);
807        break;
808    case MO_16:
809        tcg_out_ext16u(s, TCG_REG_A2, l->datalo_reg);
810        break;
811    case MO_32:
812        tcg_out_ext32u(s, TCG_REG_A2, l->datalo_reg);
813        break;
814    case MO_64:
815        tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_A2, l->datalo_reg);
816        break;
817    default:
818        g_assert_not_reached();
819        break;
820    }
821    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A3, oi);
822    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A4, (tcg_target_long)l->raddr);
823
824    tcg_out_call(s, qemu_st_helpers[size]);
825
826    return tcg_out_goto(s, l->raddr);
827}
828#else
829
830/*
831 * Alignment helpers for user-mode emulation
832 */
833
834static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addr_reg,
835                                   unsigned a_bits)
836{
837    TCGLabelQemuLdst *l = new_ldst_label(s);
838
839    l->is_ld = is_ld;
840    l->addrlo_reg = addr_reg;
841
842    /*
843     * Without micro-architecture details, we don't know which of bstrpick or
844     * andi is faster, so use bstrpick as it's not constrained by imm field
845     * width. (Not to say alignments >= 2^12 are going to happen any time
846     * soon, though)
847     */
848    tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
849
850    l->label_ptr[0] = s->code_ptr;
851    tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
852
853    l->raddr = tcg_splitwx_to_rx(s->code_ptr);
854}
855
856static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
857{
858    /* resolve label address */
859    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
860        return false;
861    }
862
863    tcg_out_mov(s, TCG_TYPE_TL, TCG_REG_A1, l->addrlo_reg);
864    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_A0, TCG_AREG0);
865
866    /* tail call, with the return address back inline. */
867    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_RA, (uintptr_t)l->raddr);
868    tcg_out_call_int(s, (const void *)(l->is_ld ? helper_unaligned_ld
869                                       : helper_unaligned_st), true);
870    return true;
871}
872
873static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
874{
875    return tcg_out_fail_alignment(s, l);
876}
877
878static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
879{
880    return tcg_out_fail_alignment(s, l);
881}
882
883#endif /* CONFIG_SOFTMMU */
884
885/*
886 * `ext32u` the address register into the temp register given,
887 * if target is 32-bit, no-op otherwise.
888 *
889 * Returns the address register ready for use with TLB addend.
890 */
891static TCGReg tcg_out_zext_addr_if_32_bit(TCGContext *s,
892                                          TCGReg addr, TCGReg tmp)
893{
894    if (TARGET_LONG_BITS == 32) {
895        tcg_out_ext32u(s, tmp, addr);
896        return tmp;
897    }
898    return addr;
899}
900
901static void tcg_out_qemu_ld_indexed(TCGContext *s, TCGReg rd, TCGReg rj,
902                                   TCGReg rk, MemOp opc, TCGType type)
903{
904    /* Byte swapping is left to middle-end expansion.  */
905    tcg_debug_assert((opc & MO_BSWAP) == 0);
906
907    switch (opc & MO_SSIZE) {
908    case MO_UB:
909        tcg_out_opc_ldx_bu(s, rd, rj, rk);
910        break;
911    case MO_SB:
912        tcg_out_opc_ldx_b(s, rd, rj, rk);
913        break;
914    case MO_UW:
915        tcg_out_opc_ldx_hu(s, rd, rj, rk);
916        break;
917    case MO_SW:
918        tcg_out_opc_ldx_h(s, rd, rj, rk);
919        break;
920    case MO_UL:
921        if (type == TCG_TYPE_I64) {
922            tcg_out_opc_ldx_wu(s, rd, rj, rk);
923            break;
924        }
925        /* fallthrough */
926    case MO_SL:
927        tcg_out_opc_ldx_w(s, rd, rj, rk);
928        break;
929    case MO_UQ:
930        tcg_out_opc_ldx_d(s, rd, rj, rk);
931        break;
932    default:
933        g_assert_not_reached();
934    }
935}
936
937static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, TCGType type)
938{
939    TCGReg addr_regl;
940    TCGReg data_regl;
941    MemOpIdx oi;
942    MemOp opc;
943#if defined(CONFIG_SOFTMMU)
944    tcg_insn_unit *label_ptr[1];
945#else
946    unsigned a_bits;
947#endif
948    TCGReg base;
949
950    data_regl = *args++;
951    addr_regl = *args++;
952    oi = *args++;
953    opc = get_memop(oi);
954
955#if defined(CONFIG_SOFTMMU)
956    tcg_out_tlb_load(s, addr_regl, oi, label_ptr, 1);
957    base = tcg_out_zext_addr_if_32_bit(s, addr_regl, TCG_REG_TMP0);
958    tcg_out_qemu_ld_indexed(s, data_regl, base, TCG_REG_TMP2, opc, type);
959    add_qemu_ldst_label(s, 1, oi, type,
960                        data_regl, addr_regl,
961                        s->code_ptr, label_ptr);
962#else
963    a_bits = get_alignment_bits(opc);
964    if (a_bits) {
965        tcg_out_test_alignment(s, true, addr_regl, a_bits);
966    }
967    base = tcg_out_zext_addr_if_32_bit(s, addr_regl, TCG_REG_TMP0);
968    TCGReg guest_base_reg = USE_GUEST_BASE ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
969    tcg_out_qemu_ld_indexed(s, data_regl, base, guest_base_reg, opc, type);
970#endif
971}
972
973static void tcg_out_qemu_st_indexed(TCGContext *s, TCGReg data,
974                                   TCGReg rj, TCGReg rk, MemOp opc)
975{
976    /* Byte swapping is left to middle-end expansion.  */
977    tcg_debug_assert((opc & MO_BSWAP) == 0);
978
979    switch (opc & MO_SIZE) {
980    case MO_8:
981        tcg_out_opc_stx_b(s, data, rj, rk);
982        break;
983    case MO_16:
984        tcg_out_opc_stx_h(s, data, rj, rk);
985        break;
986    case MO_32:
987        tcg_out_opc_stx_w(s, data, rj, rk);
988        break;
989    case MO_64:
990        tcg_out_opc_stx_d(s, data, rj, rk);
991        break;
992    default:
993        g_assert_not_reached();
994    }
995}
996
997static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args)
998{
999    TCGReg addr_regl;
1000    TCGReg data_regl;
1001    MemOpIdx oi;
1002    MemOp opc;
1003#if defined(CONFIG_SOFTMMU)
1004    tcg_insn_unit *label_ptr[1];
1005#else
1006    unsigned a_bits;
1007#endif
1008    TCGReg base;
1009
1010    data_regl = *args++;
1011    addr_regl = *args++;
1012    oi = *args++;
1013    opc = get_memop(oi);
1014
1015#if defined(CONFIG_SOFTMMU)
1016    tcg_out_tlb_load(s, addr_regl, oi, label_ptr, 0);
1017    base = tcg_out_zext_addr_if_32_bit(s, addr_regl, TCG_REG_TMP0);
1018    tcg_out_qemu_st_indexed(s, data_regl, base, TCG_REG_TMP2, opc);
1019    add_qemu_ldst_label(s, 0, oi,
1020                        0, /* type param is unused for stores */
1021                        data_regl, addr_regl,
1022                        s->code_ptr, label_ptr);
1023#else
1024    a_bits = get_alignment_bits(opc);
1025    if (a_bits) {
1026        tcg_out_test_alignment(s, false, addr_regl, a_bits);
1027    }
1028    base = tcg_out_zext_addr_if_32_bit(s, addr_regl, TCG_REG_TMP0);
1029    TCGReg guest_base_reg = USE_GUEST_BASE ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
1030    tcg_out_qemu_st_indexed(s, data_regl, base, guest_base_reg, opc);
1031#endif
1032}
1033
1034/* LoongArch uses `andi zero, zero, 0` as NOP.  */
1035#define NOP OPC_ANDI
1036static void tcg_out_nop(TCGContext *s)
1037{
1038    tcg_out32(s, NOP);
1039}
1040
1041void tb_target_set_jmp_target(uintptr_t tc_ptr, uintptr_t jmp_rx,
1042                              uintptr_t jmp_rw, uintptr_t addr)
1043{
1044    tcg_insn_unit i1, i2;
1045    ptrdiff_t upper, lower;
1046    ptrdiff_t offset = (ptrdiff_t)(addr - jmp_rx) >> 2;
1047
1048    if (offset == sextreg(offset, 0, 26)) {
1049        i1 = encode_sd10k16_insn(OPC_B, offset);
1050        i2 = NOP;
1051    } else {
1052        tcg_debug_assert(offset == sextreg(offset, 0, 36));
1053        lower = (int16_t)offset;
1054        upper = (offset - lower) >> 16;
1055
1056        i1 = encode_dsj20_insn(OPC_PCADDU18I, TCG_REG_TMP0, upper);
1057        i2 = encode_djsk16_insn(OPC_JIRL, TCG_REG_ZERO, TCG_REG_TMP0, lower);
1058    }
1059    uint64_t pair = ((uint64_t)i2 << 32) | i1;
1060    qatomic_set((uint64_t *)jmp_rw, pair);
1061    flush_idcache_range(jmp_rx, jmp_rw, 8);
1062}
1063
1064/*
1065 * Entry-points
1066 */
1067
1068static const tcg_insn_unit *tb_ret_addr;
1069
1070static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1071                       const TCGArg args[TCG_MAX_OP_ARGS],
1072                       const int const_args[TCG_MAX_OP_ARGS])
1073{
1074    TCGArg a0 = args[0];
1075    TCGArg a1 = args[1];
1076    TCGArg a2 = args[2];
1077    int c2 = const_args[2];
1078
1079    switch (opc) {
1080    case INDEX_op_exit_tb:
1081        /* Reuse the zeroing that exists for goto_ptr.  */
1082        if (a0 == 0) {
1083            tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1084        } else {
1085            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1086            tcg_out_call_int(s, tb_ret_addr, true);
1087        }
1088        break;
1089
1090    case INDEX_op_goto_tb:
1091        tcg_debug_assert(s->tb_jmp_insn_offset != NULL);
1092        /*
1093         * Ensure that patch area is 8-byte aligned so that an
1094         * atomic write can be used to patch the target address.
1095         */
1096        if ((uintptr_t)s->code_ptr & 7) {
1097            tcg_out_nop(s);
1098        }
1099        s->tb_jmp_insn_offset[a0] = tcg_current_code_size(s);
1100        /*
1101         * actual branch destination will be patched by
1102         * tb_target_set_jmp_target later
1103         */
1104        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, 0);
1105        tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1106        set_jmp_reset_offset(s, a0);
1107        break;
1108
1109    case INDEX_op_mb:
1110        tcg_out_mb(s, a0);
1111        break;
1112
1113    case INDEX_op_goto_ptr:
1114        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1115        break;
1116
1117    case INDEX_op_br:
1118        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1119                      0);
1120        tcg_out_opc_b(s, 0);
1121        break;
1122
1123    case INDEX_op_brcond_i32:
1124    case INDEX_op_brcond_i64:
1125        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1126        break;
1127
1128    case INDEX_op_ext8s_i32:
1129    case INDEX_op_ext8s_i64:
1130        tcg_out_ext8s(s, a0, a1);
1131        break;
1132
1133    case INDEX_op_ext8u_i32:
1134    case INDEX_op_ext8u_i64:
1135        tcg_out_ext8u(s, a0, a1);
1136        break;
1137
1138    case INDEX_op_ext16s_i32:
1139    case INDEX_op_ext16s_i64:
1140        tcg_out_ext16s(s, a0, a1);
1141        break;
1142
1143    case INDEX_op_ext16u_i32:
1144    case INDEX_op_ext16u_i64:
1145        tcg_out_ext16u(s, a0, a1);
1146        break;
1147
1148    case INDEX_op_ext32u_i64:
1149    case INDEX_op_extu_i32_i64:
1150        tcg_out_ext32u(s, a0, a1);
1151        break;
1152
1153    case INDEX_op_ext32s_i64:
1154    case INDEX_op_extrl_i64_i32:
1155    case INDEX_op_ext_i32_i64:
1156        tcg_out_ext32s(s, a0, a1);
1157        break;
1158
1159    case INDEX_op_extrh_i64_i32:
1160        tcg_out_opc_srai_d(s, a0, a1, 32);
1161        break;
1162
1163    case INDEX_op_not_i32:
1164    case INDEX_op_not_i64:
1165        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1166        break;
1167
1168    case INDEX_op_nor_i32:
1169    case INDEX_op_nor_i64:
1170        if (c2) {
1171            tcg_out_opc_ori(s, a0, a1, a2);
1172            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1173        } else {
1174            tcg_out_opc_nor(s, a0, a1, a2);
1175        }
1176        break;
1177
1178    case INDEX_op_andc_i32:
1179    case INDEX_op_andc_i64:
1180        if (c2) {
1181            /* guaranteed to fit due to constraint */
1182            tcg_out_opc_andi(s, a0, a1, ~a2);
1183        } else {
1184            tcg_out_opc_andn(s, a0, a1, a2);
1185        }
1186        break;
1187
1188    case INDEX_op_orc_i32:
1189    case INDEX_op_orc_i64:
1190        if (c2) {
1191            /* guaranteed to fit due to constraint */
1192            tcg_out_opc_ori(s, a0, a1, ~a2);
1193        } else {
1194            tcg_out_opc_orn(s, a0, a1, a2);
1195        }
1196        break;
1197
1198    case INDEX_op_and_i32:
1199    case INDEX_op_and_i64:
1200        if (c2) {
1201            tcg_out_opc_andi(s, a0, a1, a2);
1202        } else {
1203            tcg_out_opc_and(s, a0, a1, a2);
1204        }
1205        break;
1206
1207    case INDEX_op_or_i32:
1208    case INDEX_op_or_i64:
1209        if (c2) {
1210            tcg_out_opc_ori(s, a0, a1, a2);
1211        } else {
1212            tcg_out_opc_or(s, a0, a1, a2);
1213        }
1214        break;
1215
1216    case INDEX_op_xor_i32:
1217    case INDEX_op_xor_i64:
1218        if (c2) {
1219            tcg_out_opc_xori(s, a0, a1, a2);
1220        } else {
1221            tcg_out_opc_xor(s, a0, a1, a2);
1222        }
1223        break;
1224
1225    case INDEX_op_extract_i32:
1226        tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1227        break;
1228    case INDEX_op_extract_i64:
1229        tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1230        break;
1231
1232    case INDEX_op_deposit_i32:
1233        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1234        break;
1235    case INDEX_op_deposit_i64:
1236        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1237        break;
1238
1239    case INDEX_op_bswap16_i32:
1240    case INDEX_op_bswap16_i64:
1241        tcg_out_opc_revb_2h(s, a0, a1);
1242        if (a2 & TCG_BSWAP_OS) {
1243            tcg_out_ext16s(s, a0, a0);
1244        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1245            tcg_out_ext16u(s, a0, a0);
1246        }
1247        break;
1248
1249    case INDEX_op_bswap32_i32:
1250        /* All 32-bit values are computed sign-extended in the register.  */
1251        a2 = TCG_BSWAP_OS;
1252        /* fallthrough */
1253    case INDEX_op_bswap32_i64:
1254        tcg_out_opc_revb_2w(s, a0, a1);
1255        if (a2 & TCG_BSWAP_OS) {
1256            tcg_out_ext32s(s, a0, a0);
1257        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1258            tcg_out_ext32u(s, a0, a0);
1259        }
1260        break;
1261
1262    case INDEX_op_bswap64_i64:
1263        tcg_out_opc_revb_d(s, a0, a1);
1264        break;
1265
1266    case INDEX_op_clz_i32:
1267        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1268        break;
1269    case INDEX_op_clz_i64:
1270        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1271        break;
1272
1273    case INDEX_op_ctz_i32:
1274        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1275        break;
1276    case INDEX_op_ctz_i64:
1277        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1278        break;
1279
1280    case INDEX_op_shl_i32:
1281        if (c2) {
1282            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1283        } else {
1284            tcg_out_opc_sll_w(s, a0, a1, a2);
1285        }
1286        break;
1287    case INDEX_op_shl_i64:
1288        if (c2) {
1289            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1290        } else {
1291            tcg_out_opc_sll_d(s, a0, a1, a2);
1292        }
1293        break;
1294
1295    case INDEX_op_shr_i32:
1296        if (c2) {
1297            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1298        } else {
1299            tcg_out_opc_srl_w(s, a0, a1, a2);
1300        }
1301        break;
1302    case INDEX_op_shr_i64:
1303        if (c2) {
1304            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1305        } else {
1306            tcg_out_opc_srl_d(s, a0, a1, a2);
1307        }
1308        break;
1309
1310    case INDEX_op_sar_i32:
1311        if (c2) {
1312            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1313        } else {
1314            tcg_out_opc_sra_w(s, a0, a1, a2);
1315        }
1316        break;
1317    case INDEX_op_sar_i64:
1318        if (c2) {
1319            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1320        } else {
1321            tcg_out_opc_sra_d(s, a0, a1, a2);
1322        }
1323        break;
1324
1325    case INDEX_op_rotl_i32:
1326        /* transform into equivalent rotr/rotri */
1327        if (c2) {
1328            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1329        } else {
1330            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1331            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1332        }
1333        break;
1334    case INDEX_op_rotl_i64:
1335        /* transform into equivalent rotr/rotri */
1336        if (c2) {
1337            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1338        } else {
1339            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1340            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1341        }
1342        break;
1343
1344    case INDEX_op_rotr_i32:
1345        if (c2) {
1346            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1347        } else {
1348            tcg_out_opc_rotr_w(s, a0, a1, a2);
1349        }
1350        break;
1351    case INDEX_op_rotr_i64:
1352        if (c2) {
1353            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1354        } else {
1355            tcg_out_opc_rotr_d(s, a0, a1, a2);
1356        }
1357        break;
1358
1359    case INDEX_op_add_i32:
1360        if (c2) {
1361            tcg_out_opc_addi_w(s, a0, a1, a2);
1362        } else {
1363            tcg_out_opc_add_w(s, a0, a1, a2);
1364        }
1365        break;
1366    case INDEX_op_add_i64:
1367        if (c2) {
1368            tcg_out_opc_addi_d(s, a0, a1, a2);
1369        } else {
1370            tcg_out_opc_add_d(s, a0, a1, a2);
1371        }
1372        break;
1373
1374    case INDEX_op_sub_i32:
1375        if (c2) {
1376            tcg_out_opc_addi_w(s, a0, a1, -a2);
1377        } else {
1378            tcg_out_opc_sub_w(s, a0, a1, a2);
1379        }
1380        break;
1381    case INDEX_op_sub_i64:
1382        if (c2) {
1383            tcg_out_opc_addi_d(s, a0, a1, -a2);
1384        } else {
1385            tcg_out_opc_sub_d(s, a0, a1, a2);
1386        }
1387        break;
1388
1389    case INDEX_op_mul_i32:
1390        tcg_out_opc_mul_w(s, a0, a1, a2);
1391        break;
1392    case INDEX_op_mul_i64:
1393        tcg_out_opc_mul_d(s, a0, a1, a2);
1394        break;
1395
1396    case INDEX_op_mulsh_i32:
1397        tcg_out_opc_mulh_w(s, a0, a1, a2);
1398        break;
1399    case INDEX_op_mulsh_i64:
1400        tcg_out_opc_mulh_d(s, a0, a1, a2);
1401        break;
1402
1403    case INDEX_op_muluh_i32:
1404        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1405        break;
1406    case INDEX_op_muluh_i64:
1407        tcg_out_opc_mulh_du(s, a0, a1, a2);
1408        break;
1409
1410    case INDEX_op_div_i32:
1411        tcg_out_opc_div_w(s, a0, a1, a2);
1412        break;
1413    case INDEX_op_div_i64:
1414        tcg_out_opc_div_d(s, a0, a1, a2);
1415        break;
1416
1417    case INDEX_op_divu_i32:
1418        tcg_out_opc_div_wu(s, a0, a1, a2);
1419        break;
1420    case INDEX_op_divu_i64:
1421        tcg_out_opc_div_du(s, a0, a1, a2);
1422        break;
1423
1424    case INDEX_op_rem_i32:
1425        tcg_out_opc_mod_w(s, a0, a1, a2);
1426        break;
1427    case INDEX_op_rem_i64:
1428        tcg_out_opc_mod_d(s, a0, a1, a2);
1429        break;
1430
1431    case INDEX_op_remu_i32:
1432        tcg_out_opc_mod_wu(s, a0, a1, a2);
1433        break;
1434    case INDEX_op_remu_i64:
1435        tcg_out_opc_mod_du(s, a0, a1, a2);
1436        break;
1437
1438    case INDEX_op_setcond_i32:
1439    case INDEX_op_setcond_i64:
1440        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1441        break;
1442
1443    case INDEX_op_ld8s_i32:
1444    case INDEX_op_ld8s_i64:
1445        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1446        break;
1447    case INDEX_op_ld8u_i32:
1448    case INDEX_op_ld8u_i64:
1449        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1450        break;
1451    case INDEX_op_ld16s_i32:
1452    case INDEX_op_ld16s_i64:
1453        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1454        break;
1455    case INDEX_op_ld16u_i32:
1456    case INDEX_op_ld16u_i64:
1457        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1458        break;
1459    case INDEX_op_ld_i32:
1460    case INDEX_op_ld32s_i64:
1461        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1462        break;
1463    case INDEX_op_ld32u_i64:
1464        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1465        break;
1466    case INDEX_op_ld_i64:
1467        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1468        break;
1469
1470    case INDEX_op_st8_i32:
1471    case INDEX_op_st8_i64:
1472        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1473        break;
1474    case INDEX_op_st16_i32:
1475    case INDEX_op_st16_i64:
1476        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1477        break;
1478    case INDEX_op_st_i32:
1479    case INDEX_op_st32_i64:
1480        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1481        break;
1482    case INDEX_op_st_i64:
1483        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1484        break;
1485
1486    case INDEX_op_qemu_ld_i32:
1487        tcg_out_qemu_ld(s, args, TCG_TYPE_I32);
1488        break;
1489    case INDEX_op_qemu_ld_i64:
1490        tcg_out_qemu_ld(s, args, TCG_TYPE_I64);
1491        break;
1492    case INDEX_op_qemu_st_i32:
1493        tcg_out_qemu_st(s, args);
1494        break;
1495    case INDEX_op_qemu_st_i64:
1496        tcg_out_qemu_st(s, args);
1497        break;
1498
1499    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
1500    case INDEX_op_mov_i64:
1501    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1502    default:
1503        g_assert_not_reached();
1504    }
1505}
1506
1507static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
1508{
1509    switch (op) {
1510    case INDEX_op_goto_ptr:
1511        return C_O0_I1(r);
1512
1513    case INDEX_op_st8_i32:
1514    case INDEX_op_st8_i64:
1515    case INDEX_op_st16_i32:
1516    case INDEX_op_st16_i64:
1517    case INDEX_op_st32_i64:
1518    case INDEX_op_st_i32:
1519    case INDEX_op_st_i64:
1520        return C_O0_I2(rZ, r);
1521
1522    case INDEX_op_brcond_i32:
1523    case INDEX_op_brcond_i64:
1524        return C_O0_I2(rZ, rZ);
1525
1526    case INDEX_op_qemu_st_i32:
1527    case INDEX_op_qemu_st_i64:
1528        return C_O0_I2(LZ, L);
1529
1530    case INDEX_op_ext8s_i32:
1531    case INDEX_op_ext8s_i64:
1532    case INDEX_op_ext8u_i32:
1533    case INDEX_op_ext8u_i64:
1534    case INDEX_op_ext16s_i32:
1535    case INDEX_op_ext16s_i64:
1536    case INDEX_op_ext16u_i32:
1537    case INDEX_op_ext16u_i64:
1538    case INDEX_op_ext32s_i64:
1539    case INDEX_op_ext32u_i64:
1540    case INDEX_op_extu_i32_i64:
1541    case INDEX_op_extrl_i64_i32:
1542    case INDEX_op_extrh_i64_i32:
1543    case INDEX_op_ext_i32_i64:
1544    case INDEX_op_not_i32:
1545    case INDEX_op_not_i64:
1546    case INDEX_op_extract_i32:
1547    case INDEX_op_extract_i64:
1548    case INDEX_op_bswap16_i32:
1549    case INDEX_op_bswap16_i64:
1550    case INDEX_op_bswap32_i32:
1551    case INDEX_op_bswap32_i64:
1552    case INDEX_op_bswap64_i64:
1553    case INDEX_op_ld8s_i32:
1554    case INDEX_op_ld8s_i64:
1555    case INDEX_op_ld8u_i32:
1556    case INDEX_op_ld8u_i64:
1557    case INDEX_op_ld16s_i32:
1558    case INDEX_op_ld16s_i64:
1559    case INDEX_op_ld16u_i32:
1560    case INDEX_op_ld16u_i64:
1561    case INDEX_op_ld32s_i64:
1562    case INDEX_op_ld32u_i64:
1563    case INDEX_op_ld_i32:
1564    case INDEX_op_ld_i64:
1565        return C_O1_I1(r, r);
1566
1567    case INDEX_op_qemu_ld_i32:
1568    case INDEX_op_qemu_ld_i64:
1569        return C_O1_I1(r, L);
1570
1571    case INDEX_op_andc_i32:
1572    case INDEX_op_andc_i64:
1573    case INDEX_op_orc_i32:
1574    case INDEX_op_orc_i64:
1575        /*
1576         * LoongArch insns for these ops don't have reg-imm forms, but we
1577         * can express using andi/ori if ~constant satisfies
1578         * TCG_CT_CONST_U12.
1579         */
1580        return C_O1_I2(r, r, rC);
1581
1582    case INDEX_op_shl_i32:
1583    case INDEX_op_shl_i64:
1584    case INDEX_op_shr_i32:
1585    case INDEX_op_shr_i64:
1586    case INDEX_op_sar_i32:
1587    case INDEX_op_sar_i64:
1588    case INDEX_op_rotl_i32:
1589    case INDEX_op_rotl_i64:
1590    case INDEX_op_rotr_i32:
1591    case INDEX_op_rotr_i64:
1592        return C_O1_I2(r, r, ri);
1593
1594    case INDEX_op_add_i32:
1595    case INDEX_op_add_i64:
1596        return C_O1_I2(r, r, rI);
1597
1598    case INDEX_op_and_i32:
1599    case INDEX_op_and_i64:
1600    case INDEX_op_nor_i32:
1601    case INDEX_op_nor_i64:
1602    case INDEX_op_or_i32:
1603    case INDEX_op_or_i64:
1604    case INDEX_op_xor_i32:
1605    case INDEX_op_xor_i64:
1606        /* LoongArch reg-imm bitops have their imms ZERO-extended */
1607        return C_O1_I2(r, r, rU);
1608
1609    case INDEX_op_clz_i32:
1610    case INDEX_op_clz_i64:
1611    case INDEX_op_ctz_i32:
1612    case INDEX_op_ctz_i64:
1613        return C_O1_I2(r, r, rW);
1614
1615    case INDEX_op_setcond_i32:
1616    case INDEX_op_setcond_i64:
1617        return C_O1_I2(r, r, rZ);
1618
1619    case INDEX_op_deposit_i32:
1620    case INDEX_op_deposit_i64:
1621        /* Must deposit into the same register as input */
1622        return C_O1_I2(r, 0, rZ);
1623
1624    case INDEX_op_sub_i32:
1625    case INDEX_op_sub_i64:
1626        return C_O1_I2(r, rZ, rN);
1627
1628    case INDEX_op_mul_i32:
1629    case INDEX_op_mul_i64:
1630    case INDEX_op_mulsh_i32:
1631    case INDEX_op_mulsh_i64:
1632    case INDEX_op_muluh_i32:
1633    case INDEX_op_muluh_i64:
1634    case INDEX_op_div_i32:
1635    case INDEX_op_div_i64:
1636    case INDEX_op_divu_i32:
1637    case INDEX_op_divu_i64:
1638    case INDEX_op_rem_i32:
1639    case INDEX_op_rem_i64:
1640    case INDEX_op_remu_i32:
1641    case INDEX_op_remu_i64:
1642        return C_O1_I2(r, rZ, rZ);
1643
1644    default:
1645        g_assert_not_reached();
1646    }
1647}
1648
1649static const int tcg_target_callee_save_regs[] = {
1650    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
1651    TCG_REG_S1,
1652    TCG_REG_S2,
1653    TCG_REG_S3,
1654    TCG_REG_S4,
1655    TCG_REG_S5,
1656    TCG_REG_S6,
1657    TCG_REG_S7,
1658    TCG_REG_S8,
1659    TCG_REG_S9,
1660    TCG_REG_RA,     /* should be last for ABI compliance */
1661};
1662
1663/* Stack frame parameters.  */
1664#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
1665#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
1666#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
1667#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
1668                     + TCG_TARGET_STACK_ALIGN - 1) \
1669                    & -TCG_TARGET_STACK_ALIGN)
1670#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
1671
1672/* We're expecting to be able to use an immediate for frame allocation.  */
1673QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
1674
1675/* Generate global QEMU prologue and epilogue code */
1676static void tcg_target_qemu_prologue(TCGContext *s)
1677{
1678    int i;
1679
1680    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
1681
1682    /* TB prologue */
1683    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
1684    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
1685        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
1686                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
1687    }
1688
1689#if !defined(CONFIG_SOFTMMU)
1690    if (USE_GUEST_BASE) {
1691        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
1692        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
1693    }
1694#endif
1695
1696    /* Call generated code */
1697    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
1698    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
1699
1700    /* Return path for goto_ptr. Set return value to 0 */
1701    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
1702    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
1703
1704    /* TB epilogue */
1705    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
1706    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
1707        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
1708                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
1709    }
1710
1711    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
1712    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
1713}
1714
1715static void tcg_target_init(TCGContext *s)
1716{
1717    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
1718    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
1719
1720    tcg_target_call_clobber_regs = ALL_GENERAL_REGS;
1721    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
1722    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
1723    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
1724    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
1725    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
1726    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
1727    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
1728    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
1729    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
1730    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
1731
1732    s->reserved_regs = 0;
1733    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
1734    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
1735    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
1736    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
1737    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
1738    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
1739    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
1740}
1741
1742typedef struct {
1743    DebugFrameHeader h;
1744    uint8_t fde_def_cfa[4];
1745    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
1746} DebugFrame;
1747
1748#define ELF_HOST_MACHINE EM_LOONGARCH
1749
1750static const DebugFrame debug_frame = {
1751    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
1752    .h.cie.id = -1,
1753    .h.cie.version = 1,
1754    .h.cie.code_align = 1,
1755    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
1756    .h.cie.return_column = TCG_REG_RA,
1757
1758    /* Total FDE size does not include the "len" member.  */
1759    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
1760
1761    .fde_def_cfa = {
1762        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
1763        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
1764        (FRAME_SIZE >> 7)
1765    },
1766    .fde_reg_ofs = {
1767        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
1768        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
1769        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
1770        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
1771        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
1772        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
1773        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
1774        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
1775        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
1776        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
1777        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
1778    }
1779};
1780
1781void tcg_register_jit(const void *buf, size_t buf_size)
1782{
1783    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
1784}
1785